When I fail over two nodes while performing operations, where one of them it the 'entry point' node (i.e. the one which is the source of the streaming API), it seems the client never recovers. It does not crash or provide an exception. Operations continue failing from the workload generator.
This is using a modified Enyim (see
NCBC-121) without which I'm not even trying as it seems to crash my application during the slightest network issue.